Boundary-Aware Saliency-Based Level Set with Momentum Contrast Metaformer for Skin Cancer Segmentation and Classification

Abstract

Skin cancer is considered one of the most widespread and life-threatening cancers and remains a challenging task for dermatologists. This challenge arises due to the small boundaries and regions of affected tissue. Existing methods yield poor classification accuracy due to inefficient classifier performance in recognizing complex patterns. Therefore, an effective skin cancer segmentation and classification method called Boundary-Aware Saliency-based Level Set (BASLS) with Momentum Contrast Metaformer (MC-Metaformer) is proposed in this research. BASLS enables improved lesion segmentation by identifying significant structural components along lesion edges. Using residual connections, MC-Metaformer provides a better gradient path, addressing gradient fading issues during deep feature extraction. Preprocessing is performed on the HAM10000, ISIC-2019, and ISIC-2020 datasets to improve and standardize the data. ResNet50 is used to extract relevant features for classification. Experimental results demonstrate that the proposed MC-Metaformer outperforms the Deep Convolutional Neural Network (DCNN), achieving classification accuracies of 99.58%, 99.32%, and 98.62% on the HAM10000, ISIC-2019, and ISIC-2020 datasets, respectively. These results confirm the robustness and efficiency of the model in accurate skin cancer segmentation and classification.

ABSTRAK: Kanser kulit merupakan salah satu jenis kanser paling meluas dan mengancam nyawa serta masih menjadi cabaran utama kepada pakar dermatologi. Cabaran ini berpunca daripada sempadan dan kawasan kecil tisu yang terjejas. Kaedah sedia ada mempunyai ketepatan klasifikasi yang rendah kerana prestasi pengelasan yang kurang berkesan dalam mengenal pasti corak kompleks. Oleh itu, kajian ini mencadangkan kaedah segmentasi dan klasifikasi kanser kulit yang berkesan dikenali sebagai Boundary-Aware Saliency-based Level Set (BASLS) bersama Momentum Contrast Metaformer (MC-Metaformer). BASLS membolehkan segmentasi lesi yang lebih baik dengan mengenal pasti komponen struktur penting di sepanjang tepi lesi. Melalui penggunaan sambungan residu, MC-Metaformer menyediakan laluan kecerunan yang lebih baik bagi menangani isu kehilangan kecerunan semasa pengekstrakan ciri mendalam. Pra-pemprosesan dilakukan ke atas set data HAM10000, ISIC-2019, dan ISIC-2020 bagi meningkatkan serta menyeragam data. ResNet50 digunakan bagi mengekstrak ciri relevan untuk klasifikasi. Dapatan eksperimen menunjukkan bahawa MC-Metaformer yang dicadangkan mengatasi Rangkaian Neural Konvolusi Mendalam (DCNN) dengan ketepatan klasifikasi sebanyak 99.58%, 99.32%, dan 98.62% masing-masing pada set data HAM10000, ISIC-2019, dan ISIC-2020. Dapatan ini mengesahkan keteguhan dan kecekapan model dalam segmentasi dan klasifikasi kanser kulit secara tepat.